Open course · ~90 min
ML in the browser
A free course, in five lessons, where every concept gets touched in a real lab demo before it gets read in the text. No videos, no slides: networks training in front of you, a tokenizer learning live, a transformer's real attention. About 90 minutes total, at your own pace.
From data to your first model
Before training anything you need to understand exactly what gets trained: what goes in, what comes out, and why a model that nails every example it has already seen can fail badly on anything new.
How it actually learns
The previous lesson left a question hanging: how exactly does a model adjust its parameters? Here is the full answer, including the piece that sent the network haywire — the learning rate — explained properly.
Text that turns into numbers
Everything so far assumed the model's input was already numbers. Text is not. This lesson covers how a word turns into something a network can multiply — and why that conversion partly decides what the model can ever understand.
The transformer, no smoke and mirrors
Tokens and embeddings give you a sequence of vectors. What is missing is the piece that decides what to do with that sequence: the attention mechanism, the heart of every current language model, explained with the real weights of a small transformer you can poke at.
Models in the real world
A model that works in a demo is not finished: it still needs to fit wherever it has to run, to know when to doubt itself, and to be evaluated honestly. Final lesson, and the most practical one.